Technology portfolios rarely kill anything. Projects are approved with energy, launched with announcements and then allowed to continue because stopping them would mean admitting something. The sponsor would lose face. The team would need new work. The vendor contract has eighteen months to run. Someone on the executive committee likes it.
AI makes this harder. Many AI use cases are cheap to keep running at small scale, so there is never a single large bill that forces the question. Usage often stays respectable even when value is unproven, because staff have been told to use the tool. And AI carries an aura of inevitability: stopping an AI initiative can feel like falling behind, even when the specific initiative is not working.
CloudNala's position is that stopping is a governance outcome, not a failure, and the criteria for it must be written down before the use case is funded. When stop criteria are agreed in advance, a stop decision is a rule being applied. When they are invented at the time, it is a sponsor being overruled, and that almost never happens.
Indigo Mutual is fictional, and its systems and figures are illustrative. Its board now sees a one-page value view with forecast, evidenced and verified benefit in separate columns, as described in Part 5. Every row ends in a decision. This part is about how those decisions are made.
Four outcomes, not two
A funding review is usually framed as continue or stop. That framing is part of the problem, because "stop" is so drastic that almost everything gets "continue". Indigo Mutual uses four outcomes:
- Scale: the evidence supports more investment, wider rollout or a larger scope.
- Hold: continue at the current level for a fixed period while a specific gap is closed, usually missing evidence.
- Redesign: the problem is still worth solving but this approach is not working; change the design before spending more.
- Stop: end funding, and hand the use case to a structured retirement.
The middle two options make the outer two easier. A team facing "redesign" can keep its problem and its sponsor while admitting the first attempt was wrong. "Hold" makes a missing-evidence situation explicit and time-limited rather than indefinite.
The decision sequence
Indigo Mutual's committee works through the same three questions for every material use case each quarter, in the same order.
Is there an accountable owner, named and active? This comes first because without an owner, none of the other answers can be acted on. A use case whose sponsor has left and whose benefit nobody has accepted is stopped, regardless of how promising it looks. If someone wants it to continue, they can take ownership, and then it comes back next quarter.
Is the benefit evidenced against the baseline? If not, the use case goes on hold with a fixed deadline, normally ninety days, to instrument it properly. At the end of the hold, it either has evidence or it is treated as having none. The hold is not renewable without a committee decision recorded in the minutes.
Are unit cost and controls within the agreed bounds? A use case can deliver real benefit at a cost per outcome that makes no sense, or while breaching its risk boundary. Either sends it to redesign before any more money goes in.
Only a use case that passes all three is a candidate to scale, and it scales with the same measures in place so the next review can tell whether scaling worked.
Six triggers for a stop review
Between quarterly reviews, certain conditions trigger an early look. Indigo Mutual wrote these down at approval for each use case, with thresholds that fit the use case.
Weak adoption. If fewer than an agreed share of eligible users use the tool after two quarters, the tool probably does not fit the work. That is a redesign question first, and a stop question if redesign does not change it.
Unverified benefit. If, after twelve months, nothing has moved beyond the forecast level, the use case has had time to prove itself and has not.
High exception rate. If more cases are sent back to people than the AI handles cleanly, it may be creating work rather than saving it. Part 3's review-step instrumentation is what makes this visible.
Rising unit cost. If cost per good outcome rises for two reviews running, something is drifting: volume of difficult cases, model pricing, retries or rework.
Control failures. Any breach of the risk boundary agreed at approval triggers an immediate review, and depending on severity, an immediate pause.
No accountable owner. A sponsor leaving without a successor accepting the benefit is a stop trigger in its own right.
The thresholds in the diagram are examples, not norms. A customer-facing credit assistant at a lender should have much tighter control thresholds than an internal drafting tool. What matters is that the thresholds exist before the review, so nobody is negotiating them in the meeting.
Why "hold" needs a clock
Of the four outcomes, hold is the one most easily abused. It feels responsible: we are not stopping, we are gathering evidence. In practice, without a deadline, hold becomes a permanent state where use cases nobody wants to defend or kill can live indefinitely.
Indigo Mutual's rule is that a hold has a single stated reason, a single deliverable that resolves it and a date. "Hold pending evidence of handling-time reduction, journey instrumentation live by 31 January, reviewed in February" is a hold. "Hold pending further analysis" is not, and the committee secretary is asked to reject it.
When the staff productivity assistant from Part 4 came up for review, it went on hold with exactly that kind of clock. The ninety days were used to sample work output from teams with and without heavy use. The evidence was mixed: real time savings in some roles, none visible in others. The outcome was a redesign: licences were concentrated in the roles where the evidence was positive, and 1,400 seats were not renewed.
Stopping well
Deciding to stop is half the job. Stopping safely is the other half, and it is easy to get wrong with AI. Users may have built workarounds that depend on the tool. Data may have been stored in places that need to be cleaned up. Downstream processes may consume its output. Contracts may have notice periods.
Retirement is a governed process of its own, covering the inventory of users and dependencies, data retention and deletion, communications and support end dates. Archived does not mean failed, part of the From Experiment to Trusted Platform series, covers how to decommission well. The funding decision in this part should trigger that process, not replace it.
It also helps to record what was learnt. A stopped AI use case at Indigo Mutual produces a one-page note: what it was for, what the evidence showed, why it stopped, and what would need to be true for a future attempt to succeed. Those notes are surprisingly valuable when the same idea returns two years later with a new vendor.
The point of stop criteria
Organisations sometimes resist stop criteria because they fear being too harsh on young use cases. The opposite is usually true. Clear criteria protect promising use cases from being cut arbitrarily in a budget squeeze, because they can point to evidence that they meet the bar. They also release money and attention from the ones that do not, which is the only way a portfolio improves.
A board that never stops anything is not supporting innovation. It is funding everything at the level of its weakest idea.
Next: Part 7, Vendor promise or competitive advantage?. All parts are on the series page.
How CloudNala can help
We help executive committees define scale, hold, redesign and stop criteria for their AI portfolio, set use-case-specific thresholds at approval, and run the first quarterly review with them. Where a use case is stopped, we help plan its retirement so users, data and dependencies are handled properly.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an AI Value Review or write to us at consult@cloudnala.co.za